What is it about?
AI agents depend on the information they are given to do their work, but there has been little agreement about what makes that information “good.” This paper introduces CAFE(S), a framework for defining high-quality AI context across five dimensions: Clarity, Actionability, Fidelity, Efficiency, and Security. It shows how problems such as ambiguous instructions, missing goals, outdated information, irrelevant details, and unsafe inputs can cause agents to fail even when the underlying AI model is capable. CAFE(S) provides a shared language for recognizing these problems and practical guidance for improving the context we provide to AI agents.
Featured Image
Photo by Jayden Sim on Unsplash
Why is it important?
As we delegate increasingly complex tasks to agents, the consequences of poor context grow. It can waste time and money when agents head in the wrong direction or need repeated correction. It can frustrate developers who have to spend more time guiding agents, checking their work, and fixing mistakes. And as we trust agents with more important work, poor context can lead to security, compliance, and even legal risks. Improving context quality can help make AI agents more effective, less costly, and safer to use.
Perspectives
I've spent much of my career studying what helps software developers do their best work. As agents become a central part of how we build software, I've become increasingly interested in what helps humans and agents do their best work together. CAFE(S) grew out of thinking about that question. I'm excited about this work not because I think we have all the answers, but because I hope we've given researchers and practitioners a shared language for asking better questions about context and human-agent collaboration.
Brian Houck
DX
Read the Original
This page is a summary of: CAFE(S): Your Agent Is Only As Good As Its Context, Queue, August 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3847288.
You can read the full text:
Contributors
The following have contributed to this page
